Port real AI models, run them on actual edge hardware, and share reproducible results. Pick a quest (Vision, Speech, LLMs, Robotics, Systems, or Wildcard), join the Hugging Face org, and submit your work via the CORE-ET model repo. Beginners start with existing benchmarks; experts push new kernels and novel models.

🚀 What Is the AIFoundry CORE-ET Hackathon?

The AIFoundry CORE-ET Hackathon is a community-driven event where participants take open-source AI models and make them run efficiently on real edge hardware — think Raspberry Pi, RISC-V boards, embedded devices, and other low-power silicon.

The mission is simple: Port real models. Run them on real silicon. Share the results.

This isn’t a cloud-only hackathon. The goal is to bring AI workloads closer to the edge — where latency, power, and privacy matter most.

🔗 Hackathon Hub: huggingface.co/AIFoundry-hackathon

🧩 The Six Quests You Can Choose From

Quest	What You Can Build
👁️ Vision	Denoising, detection, camera-first edge workloads
🎙️ Speech + Audio	Whisper-style kernels and streaming-friendly pipelines
🧠 Local LLMs	GGUF and llama.cpp flows on ET-backed runtimes
🤖 Robotics + Drones	Perception, control, and low-power deployment paths
⚙️ Systems	Faster kernels, better memory movement, cleaner board runners
🧪 Wildcard	Bring your own open model and make the benchmark reproducible

🧭 How Participation Works (The Universal Flow)

Whether you’re a beginner or an expert, the core workflow is the same:

1. 🤝 Join the Hugging Face organization
2. 👯 Team up or go solo
3. 🎯 Pick a model, kernel, benchmark, or demo target
4. 📚 Use the CORE-ET model repo to follow the technical path
5. 🚀 Submit your port or improvement
6. 📊 The board workflow runs and comments back with results

🔗 Join the Org: huggingface.co/AIFoundry-hackathon
🔗 CORE-ET Model Repo: Link on the org page
💬 Discord Help: discord.gg/CbSA2umxf6

🌱 PATH A: Beginner’s Roadmap

“I know some Python, I’ve heard of AI, but I’ve never ported a model to hardware before.”

Step 1: Set Up Your Accounts (5 minutes)

What	Why	Link
Hugging Face Account	Where models live and where the hackathon happens	huggingface.co/join
GitHub Account	Where code lives and where you submit	github.com/signup
Discord Account	Where you ask for help	discord.com

Action: Create all three. Join the AIFoundry Hackathon Discord and say hi in #community-lab.

Step 2: Join the Organization (2 minutes)

1. Go to huggingface.co/AIFoundry-hackathon
2. Click “Request to Join”
3. Wait for approval (usually quick)

Why this matters: The org page is your public roster — teams, models, demos, and board results all live here.

Step 3: Pick the Easiest Quest — Vision (5 minutes)

For beginners, Vision is the friendliest starting point. Why?

• Lots of existing examples (DnCNN, YOLO)
• You can see results visually (images in, images out)
• Hugging Face has tons of pre-trained vision models

Action: Commit to the Vision quest. Write it down.

Step 4: Explore the CORE-ET Model Repo (15 minutes)

1. From the org page, find and click the CORE-ET model repo
2. Look for folders like:

• examples/ — starter code
• benchmarks/ — how performance is measured
• docs/ — instructions
• vision/ — your quest area

Action: Read the README. Don’t understand everything? That’s fine. Just get familiar with the layout.

Step 5: Reproduce an Existing Benchmark (Your First Win!)

This is the most important step for beginners. Don’t build new — copy first.

What to do:

1. Find an existing benchmark in the repo (e.g., DnCNN denoising or YOLO detection)
2. Read the README.md inside that folder
3. Follow the setup instructions EXACTLY
4. Run the code on your computer (cloud is fine for now!)

Why this matters:

• You learn the submission format
• You understand the board workflow
• You prove the system works before you change anything

Action: Pick ONE existing benchmark and get it running. Celebrate when it works!

Step 6: Make One Small Improvement (Your First Contribution)

Once the benchmark runs, make a tiny change:

Easy Win	What to Do
Swap the image	Use a different test image and show it still works
Add logging	Print “Model loaded successfully” or timing info
Try a different model	Replace microsoft/resnet-50 with google/vit-base-patch16-224
Document better	Rewrite the README with clearer steps for the next beginner

Action: Make one change, test it, and document what you did.

Step 7: Submit Your First PR (Pull Request)

1. Fork the CORE-ET repo on GitHub (creates your copy)
2. Clone it to your computer: git clone <your-fork-url>
3. Create a branch: git checkout -b my-first-vision-improvement
4. Add your changes (code + documentation)
5. Commit: git commit -m "Improved DnCNN benchmark with clearer docs"
6. Push: git push origin my-first-vision-improvement
7. Open a Pull Request on GitHub

The board workflow will run automatically and comment with results.

🎯 Beginner’s Milestones

Milestone	What It Looks Like
✅ Week 1	Joined org, read docs, reproduced one benchmark
✅ Week 2	Made first small improvement, submitted first PR
✅ Week 3	Helped another participant in Discord
✅ Week 4	Submitted a complete demo showing your model running

🚀 PATH B: Expert’s Roadmap

“I know PyTorch, I’ve optimized models before, and I want to push hardware limits.”

Strategy 1: Pick a Hard Quest — Systems or Wildcard

Quest	Expert Angle
⚙️ Systems	Write faster kernels, optimize memory movement, create cleaner board runners
🧪 Wildcard	Port a cutting-edge model no one has tried on edge hardware yet
🤖 Robotics + Drones	Real-time perception pipelines with strict latency budgets

Action: Choose based on your expertise. Systems experts go for kernels. Model experts go for Wildcard.

Strategy 2: Deep Dive into the Board & Runtime

Experts need to understand:

• What silicon is supported? (RISC-V? ARM? Custom?)
• What runtime? (ExecuTorch? llama.cpp? Custom?)
• What benchmark metrics matter? (Latency? Throughput? Power? Accuracy?)

Action: Read the technical docs in the CORE-ET repo thoroughly. Join Discord and ask the maintainers about hardware access.

Strategy 3: Port a Novel Model

Instead of improving existing work, bring something new:

1. Pick a recent open model from Hugging Face Hub
2. Pin the exact repo, revision, and file references
3. Write the port layer (model → runtime → board)
4. Create reproducible benchmark scripts
5. Document board results with numbers

Key requirement: Your submission must be reproducible. Anyone should be able to run your code and get the same results.

Strategy 4: Optimize at the Kernel Level

For Systems quest participants:

• Profile existing kernels to find bottlenecks
• Rewrite hot paths in optimized C/assembly
• Improve memory layout and data movement
• Document before/after performance on actual hardware

Action: Find the slowest kernel in an existing benchmark. Make it 2x faster. Submit with proof.

Strategy 5: Mentor & Build Community

Experts can multiply their impact:

• Answer questions in Discord #community-lab
• Review PRs from beginners
• Write tutorials for complex ports
• Pair with beginners to help them reproduce your work

Action: Help one beginner reproduce a run. It’s listed as a “good first contribution” for a reason.

🎯 Expert’s Milestones

Milestone	What It Looks Like
✅ Day 1-2	Understood board runtime, identified optimization target
✅ Week 1	Submitted first kernel improvement or novel model port
✅ Week 2	Achieved measurable speedup or new model running on hardware
✅ Week 3	Mentored a beginner, improved documentation
✅ Week 4	Submitted a demo that shows real-world edge deployment

📊 Good First Contributions (For Everyone)

Contribution	Beginner?	Expert?	Impact
Improve DnCNN benchmark	✅ Easy	✅ Can optimize further	Medium
Improve YOLO detection	✅ Moderate	✅ Can rewrite kernels	High
Add new open model with pinned HF reference	❌ Hard	✅ Core strength	Very High
Make a kernel faster + document board result	❌ Very Hard	✅ Core strength	Very High
Package a small demo	✅ Easy	✅ Can make it production-ready	Medium
Help another participant reproduce a run	✅ Easy	✅ Easy	High (community!)

💬 Getting Help

Channel	Best For
Discord #community-lab	Quick questions, finding teammates, debugging
GitHub Issues	Bug reports, feature requests, technical discussions
Hugging Face Org Discussions	General hackathon updates, announcements

When asking for help, include:

• What quest you’re working on
• What hardware (if any) you have access to
• What you’ve tried so far
• The exact error message (if stuck)

🏁 Final Checklist

Before You Start:

• [ ] Created Hugging Face account
• [ ] Created GitHub account
• [ ] Joined AIFoundry Hackathon org
• [ ] Joined Discord
• [ ] Read the CORE-ET model repo README
• [ ] Picked a quest

During the Hackathon:

• [ ] Reproduced at least one existing benchmark
• [ ] Made and submitted one improvement
• [ ] Documented your work clearly
• [ ] Engaged with the community (asked or answered a question)

Before Submitting:

• [ ] Code runs without errors
• [ ] Results are reproducible (pinned model references)
• [ ] README explains what you did and why it matters
• [ ] Board workflow passes (or you understand why it fails)

Remember: This hackathon is about making AI real on real hardware. Whether you’re writing your first Python script or hand-optimizing assembly kernels, there’s a place for you. Start where you are, use what you have, and help someone else along the way.

Happy hacking! 🤗⚡

Originally published on AskGenAI. For questions or updates, join the AIFoundry Discord.